How to coordinate at best a set of autonomous software entities that need to achieve a collective goal, without centralised supervision? This is the question that our research in adaptive, self-organising, decentralised coordination mechanisms seeks to answer.
Motivation
Agents in multi-agent systems are usually equipped with pre-defined interaction means (e.g. messaging abilities) and fixed coordination protocols to abide to. This cannot cope with highly dynamic scenarios demanding for adaptation.
A way out is to let agents learn how to interact and coordinate at best.
Coordination in pervasive systems cannot be done by individually and imperatively programming each partecipating device: the levels of abstraction and autonomy are too low.
Ways to let devices autonomously figure out how to participate in a systemic goal given by designer, or arising dynamically according to context, must be found.
Approach
We deal with these open challenges by using techniques from
- multi-agent reinforcement learning
- causal reasoning
- computational argumentation
Reference publications
- Developing a “Sense of Agency” in IoT Systems: Preliminary Experiments in a Smart Home Scenario
- Argumentation-Based Coordination in IoT: A Speaking Objects Proof-of-Concept
- Distributed Discovery of Causal Networks in Pervasive Environments
- Learning Stigmergic Communication for Self-organising Coordination
- Cooperative Driving at Intersections Through Agent-Based Argumentation